Service Overview

Physical AI

FPT Automotive's physical-ai service implements and validates advanced embodied AI for autonomous driving and robotics. We support OEM roadmaps through structured robot manipulation data collection (Phase 0 RealSense, 250 episodes), a Success Rate evaluation framework (Total, Grasp, Transport, Place SR) to benchmark π0.5 and GR00T under identical conditions, and predictive safety models. Powered by an event-driven Vertex AI pipeline with 100% Model Garden integration and BigQuery logs.

Autonomous DrivingRoboticsMachine Learning

250

robot episodes collected

100%

Model Garden integration

3-15s

CAN hazard pre-warning

<2

false alarms per 1,000km

Dual

Onsite-offshore robot ops

Capabilities

Key capabilities

Multi-Modal Perception

Fusion of vision, LiDAR, and radar for comprehensive environmental understanding and object detection.

Real-Time Scene Understanding

Semantic segmentation and scene interpretation enabling intelligent decision-making in complex environments.

Autonomous Navigation

Intelligent path planning and obstacle avoidance for autonomous vehicles and mobile robots.

Robotic Manipulation

Precise control and decision-making for robotic arms and manufacturing systems.

Edge-Based Inference

Sub-100ms latency inference enabling real-time autonomous operation without cloud dependency.

Technology

Technology stack

Physical AI architecture diagram
Component Technology
Vision VLM, Computer Vision
Sensor Fusion LiDAR, Radar, Camera
Robotics ROS, Motion Planning
Edge Computing NVIDIA Orin, Qualcomm
Development Python, C++, CUDA

Use cases

Real-world applications

Documented outcomes from actual deployments.

1

Robot Manipulation Data Collection & Benchmarking

End-to-end data collection protocol for a 12-DoF robot arm using 4 Intel RealSense D405 cameras. Structured Success Rate framework (Total, Grasp, Transport, Place SR) with N=30 trials per set — establishing the baseline to benchmark π0.5 and GR00T models.

Before

No rigorous data protocol, no standardized evaluation framework; earlier effort yielded <70% inference accuracy with 1,500 episodes

After

250 high-quality episodes collected with a reproducible SR evaluation framework and phase-separated requirements

250 episodes collected
4 RealSense cameras
30 trials per eval set
2

Event-Driven Vertex AI Training Pipeline

Eventarc-triggered Vertex AI pipeline with 4 sequential steps and 100% Model Garden integration. BigQuery-backed logs split into run-level and step-level categories for immediate diagnostics. Cross-cloud ingress from AWS S3 and GCS.

Before

Pipeline visibility fragmented across AWS S3 and GCP; reproducibility and root-cause analysis slow

After

4-step pipeline auto-executes from data upload trigger; all logs SQL-filterable in BigQuery

4 steps auto-executing
100% Model Garden integration
2 BigQuery log categories

How we work

Implementation approach

1

Phase 1: Perception System Design

  • Define sensor configuration and placement
  • Design multi-modal fusion architecture
  • Plan edge computing infrastructure
2

Phase 2: Model Development

  • Develop VLM models for scene understanding
  • Train object detection and segmentation models
  • Optimize models for edge hardware
3

Phase 3: Integration & Testing

  • Integrate with vehicle/robot control systems
  • Conduct real-world testing in target environments
  • Validate safety and performance metrics
4

Phase 4: Deployment & Optimization

  • Deploy to autonomous vehicle/robot fleet
  • Monitor performance and collect data
  • Continuously improve models based on real-world data

Explore more

Related services